cdeust/Cortex

v3.13.0 Feature

This release adds 2 notable features for engineering teams evaluating rollout.

Published 1mo MCP Data & Storage

View tool

✓ No known CVEs patched

Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

agent-memory-system anthropic artificial-intelligence causal-inference claude claude-code

+14 more

claude-code-plugin cognitive-architecture cognitive-science episodic-memory hopfield-network llm-memory long-term-memory mcp-server model-context-protocol neuroscience persistent-memory predictive-coding retrieval-augmented-generation vector-db

Summary

AI summary

Marketplace now supports no‑Python installation with uvx‑consistent hooks and adds the cortex-doctor CLI.

Full changelog

feat(marketplace): no-Python-required install + uvx-consistent hooks
bench(phase-5): LongMemEval 500-Q regression gate PASSED — exact match
release(v3.13.0): bump version + cortex-doctor CLI for marketplace users
feat(phase-7): hardening — content normalization, metrics, Dockerfile
feat(phase-4): chunked consolidate + streaming homeostatic moments
feat(phase-2): JOIN replacements for plasticity / synaptic tagging / co-activation
test(phase-5-step-6): I10 invariant — pool capacity vs cycle worker fanout
feat(phase-5-step-4c): asyncio.to_thread wrap handlers at registration
feat(phase-5-step-4b): migrate 14 external store._conn.execute sites to pool
feat(phase-5-step-4a): _execute borrows from interactive_pool per call
feat(phase-5-step-5): per-tool admission semaphore middleware
feat(phase-5-step-3): latency-class registry for MCP tool handlers
feat(phase-5-step-2): ConnectionPool infrastructure in PgMemoryStore
feat(phase-5-step-1): psycopg_pool dep + pool config knobs
docs(phase-5): ConnectionPool + to_thread + admission design spec
docs(darval): github issue #14 A3 delivery reply draft
bench(a3-step-9): BEAM-100K regression gate PASSED — exact zero delta
bench(a3-step-9): LoCoMo 1982-Q regression gate PASSED
bench(a3-step-9): LongMemEval 500-Q regression gate PASSED
fix(a3): effective_heat underflow guards for multi-year age rows
feat(a3-step-8): delete legacy heat path — single A3 canonical implementation
feat(a3-step-7): flag-gated decay cycle no-op under A3_LAZY_HEAT
feat(a3-step-6): recall_memories_lazy PL/pgSQL function + get_a3_ddl
feat(a3-step-5): dual-path heat writers (anchor, preemptive_context, stale, sqlite)
chore(settings): add project permission allowlist for read-only ops
style: ruff format + drop unused import in Phase 6 fixes
feat(reranker): Platt calibration from rate_memory feedback (AF-2)
feat(a3-step-4): bump_heat_raw canonical writer + homeostatic_state helpers
feat(write_gate): auto-calibrate threshold from observed acceptance rate (AF-5)
feat(a3-step-2): effective_heat() + effective_heat_frozen() PL/pgSQL functions
feat(a3-step-1): CORTEX_MEMORY_A3_LAZY_HEAT flag + migration SQL scaffolding
fix(recall_hierarchical): require domain or memory_ids — kill O(N^2) uncapped fallback
style: ruff check --fix unused imports in Phase 1 tests
style: ruff format Phase 1 changes — line length + minor normalisation
fix(profile-store): per-domain split — bound write amplification per session
fix(compression): eliminate redundant gist encode on 0→2 transitions
fix(homeostatic): Welford one-pass moments for heat-distribution health
fix(wiki-sync): surface errors in remember handler, stop silent swallowing
fix(codebase-analyze): bounded-candidate rglob — ADR-0045 R2/R3
style: ruff format v3.12.x commits (CI format-check fix)
fix(validation): bounded tags envelope on remember (ADR-0045 R2)
fix(validation): tighten remember.content maxLength 50K→10K (ADR-0045 R2)
fix(embedding_cache): SHA256[:16] keys, never raw text (ADR-0045 R5)
fix(import_sessions): delete full_read path — streaming only (ADR-0045 R2)